CN103985132A - Mismatching point iterative detection method based on K neighbor graphs - Google Patents

Abstract

The invention provides a mismatching point iterative detection method based on K neighbor graphs. The method overcomes defects in an existing mismatching point detection method and improves the detection precision of mismatching points. The method comprises the steps of (1) dividing matching points into a plurality of sets, and performing mismatching point detection on each set of matching points respectively according to the step (2) to step (5); (2) building the K neighbor graph on an original image; (3) for each pair of matching points, obtaining K pairs of matching points with the closest distance through the K neighbor graph, building a transformational model by using the K pairs of matching points, and taking the error between the matching points and the transformational model as the error of the matching points; (4) deleting the S matching points with the largest error; (5) performing iterative detection to end judgment. If the error of all the remaining matching points is smaller than a given threshold value, iteration is stopped, otherwise, N=N-S, and the step (2) is executed.

Description

Mismatching point iteration detection method based on k nearest neighbor figure

Technical field

The present invention relates to image processing techniques, specifically, is a kind of Mismatching point iteration detection method based on k nearest neighbor figure, for detection of and delete the Mismatching point producing in remote sensing images couplings, improve remote sensing images matching precision.

Background technology

Remote sensing images coupling is an important step in remote sensing image processing, is widely used in fields such as remote sensing image registration, target identification, target following, variation detection, Image Mosaics.Remote sensing images coupling is the high precision match point of finding between two width images by certain matching algorithm.In remote sensing images coupling, remote sensing images to be matched are called original image, are called benchmark image as the image of coupling benchmark.In recent years, along with the development of image matching technology, matching precision and automatization level improved constantly, and the especially yardstick taking SIFT as representative, invariable rotary Feature Correspondence Algorithm makes the precision of remote sensing images coupling obtain great lifting.But, due to the impact of the factors such as illumination, feature changes, different imaging angle and imaging time, geometry deformation, no matter adopt which kind of image matching algorithm, always produce Mismatching point.Therefore, for remote sensing images coupling, it is a necessary step that Mismatching point detects, and its precision directly affects the precision of images match.

In remote sensing images coupling, stochastic sampling coherence method (Random Sample Consensus, RANSAC) and polynomial fitting method are generally used for Mismatching point and detect.RANSAC method is to concentrate from one group of sample data that comprises abnormal data, the alternative manner of estimation optimization model parameter (transformation models between two given width images).In Mismatching point detects, conventionally use fundamental matrix or homography matrix as transformation model, and the larger match point of transformation model error is exactly Mismatching point.Polynomial fitting method is that all correct match points of hypothesis meet some multinomial models, then use the methods such as least square fitting to solve polynomial coefficient, finally calculate the error of each match point and multinomial model, the match point that error is greater than setting threshold is exactly Mismatching point.Although principle, the computing method of RNASAC method and polynomial fitting method are different, be all that all correct match points of hypothesis meet a unified model.But for the larger remote sensing images of part coverage, be subject to the impact of complex-terrain, geometry deformation etc., all correct match points cannot meet same model, and therefore, conventional RANSAC method and polynomial fitting method cannot be directly used in the mistake matching detection of such remote sensing images.

Summary of the invention

The present invention is directed to the deficiency of existing Mismatching point detection technique, proposed a kind of Mismatching point iteration detection method based on k nearest neighbor figure, complete accurately Mismatching point and detect, further improve the precision of images match.Prerequisite of the invention process is to use image matching method to obtain match point.

Technical scheme of the present invention is as follows:

A Mismatching point iteration detection method based on k nearest neighbor figure, is characterized in that comprising the following steps:

(1) match point grouping.To match point, according to the coordinate of match point on original image, match point is divided into some groups for N to be detected, then each group match point is carried out to Mismatching point detection according to step (2)-(5) respectively;

(2) on original image, build k nearest neighbor figure.A summit using each match point in original graph as k nearest neighbor figure, for each summit, if the k nearest neighbor point on Shi Gai summit, other summit, there is a nonoriented edge in these two summits; Nonoriented edge between summit and summit has formed k nearest neighbor figure.

(3) calculate the error of every a pair of match point.For every a pair of match point, obtain nearest K to match point by k nearest neighbor figure, then use K to set up transformation model to match point, calculate the error of every pair of match point and transformation model as this error to match point.

(4) match point of deletion error maximum.After the error of all match points has been calculated, delete the S of error maximum to match point;

(5) iterative detection stops judgement.If the error of all residue match points is all less than given threshold value, stop iteration; Otherwise N=N-S, returns to step (2).

Further, the match point grouping in step (1), the method that adopts grid to divide, first carries out grid division to original image, and sizing grid is M*N, then according to the coordinate of match point on original image, all match points is assigned to different grids; The value of M and N is determined according to original image size, ensures that each grid has up to a hundred match points.

Further, in step (2), the value of K is determined according to the transformation model of step (3) selection, and the value of K is greater than the needed minimum match point number of computational transformation model parameter.

Further, the k nearest neighbor figure in step (2) is non-directed graph, uses adjacency list storage, is designated as G=(V, E), the set that wherein V is summit, the set that E is limit; Structure criterion is as follows: a summit using each match point in original image as k nearest neighbor figure, forms vertex set V=v ₁... v _n, N is total number of match point in original image; For vertex v _iif, vertex v _jv _ik nearest-neighbor in point, v _iand v _jbetween there is a nonoriented edge; Vertex v _ik nearest-neighbor in point be apart from vertex v _ia nearest K summit; Distance between two summits is the Euclidean distance of match point on original image corresponding to summit.

Further, the transformation model in step (3) is n rank polynomial expression, fundamental matrix, homography matrix.

Further, the match point error calculation formula in step (3) is:

$\mathrm{error}=\sqrt{{(x_i-x_i^{\text{'}})}^2+{(y_i-y_i^{\text{'}})}^2}$

In above formula, (x _i, y _i) be matching double points M _icoordinate on original image, (x ' _i, y ' _i) be matching double points M _icoordinate on benchmark image, is converted to the coordinate on original image according to the transformation model building.

The present invention compared with prior art has following advantages: by building k nearest neighbor figure, only realized and set up transformation model at image local.Because image local area scope is less, therefore, all correct match points can meet same transformation model, thereby have well overcome the shortcoming of traditional RANSAC method and polynomial fitting method, have improved the precision that Mismatching point detects.

Brief description of the drawings

The Mismatching point iterative detection process flow diagram of Fig. 1 based on k nearest neighbor figure

Embodiment

Now by reference to the accompanying drawings, a kind of embodiment of the present invention is described.

The Mismatching point iterative detection process flow diagram of Fig. 1 based on k nearest neighbor figure, comprises five steps:

(1) match point grouping.For the remote sensing image of big data quantity, because the match point obtaining is more, if directly carry out Mismatching point detection, can cause speed slow, therefore, the match point first images match being obtained divides into groups.The method that match point grouping adopts grid to divide, first carries out grid division to original image, and sizing grid is M*N, then according to the coordinate of match point on original image, all match points is assigned to different grids.The value of M and N is determined according to original image size, ensures that each grid has up to a hundred match points.The match point of different grids forms some groups of match points.For each group match point, carry out Mismatching point detection according to step (2)-(5) respectively.

(2) on original image, build k nearest neighbor figure.Figure is that between data, the one of relation represents mode, is made up of the set on summit and the set on limit (relation between summit).If limit does not have direction, be called non-directed graph, if there is direction on limit, be called digraph.In patent of the present invention, the k nearest neighbor figure of structure is non-directed graph, is designated as G=(V, E), the set that wherein V is summit, the set that E is limit.Wherein the value of K is determined according to the transformation model of step (3) selection, and the value of K is greater than the needed minimum match point number of computational transformation model parameter.Concrete structure criterion is as follows:

A summit using each match point in original image as k nearest neighbor figure, forms vertex set V=v ₁... v _n, N is total number of match point in original image.The initial value of N is the total number of match point that automatic matching method obtains, and along with continuous iterative detection and deletion, the value of N can constantly reduce.For vertex v _iif, vertex v _iv _ik nearest-neighbor in point, v _iand v _ibetween there is a nonoriented edge.Vertex v _ik nearest-neighbor in point be apart from vertex v _ia nearest K summit.Distance between two summits is the Euclidean distance of match point on original image corresponding to summit.

The storage means of figure comprises adjacency matrix and adjacency list.Adjacency matrix is the matrix of a n*n, stores the information on limit in matrix.And adjacency list is another kind of storage organization, each vertex v _iall of its neighbor point form a linear list, its advantage is to obtain easily K the abutment points on summit.In patent of the present invention, need to obtain the nearest K of a certain match point of distance to match point by k nearest neighbor figure, therefore adopt the storage means of adjacency list.

(3) calculate the error of every a pair of match point.Build after k nearest neighbor figure, for each match point, use the match point of its K arest neighbors to calculate its error with respect to setting models.For every a pair of match point M _i, i=1 ..., N first according to the match point on original image, obtains K to match point from k nearest neighbor figure, then uses K to set up given transformation model to match point.In remote sensing images, conventionally use the transformation models such as n rank polynomial expression, fundamental matrix, homography matrix.Can select any one transformation model to carry out error calculating.In an embodiment of the present invention, use single order multinomial model, polynomial coefficients by using least-square fitting approach calculates and obtains; Finally calculate match point M _iwith the error of setting models, and using this error as match point M _ierror.Error calculation formula is as follows:

$\mathrm{error}=\sqrt{{(x_i-x_i^{\text{'}})}^2+{(y_i-y_i^{\text{'}})}^2}$

In above formula, (x _i, y _i) be matching double points M _icoordinate on original image, (x ' _i, y ' _i) be matching double points M _ifor the coordinate on benchmark image, be converted to the coordinate on original image according to the transformation model building.

(4) S that deletes error maximum is to match point.After the error of all match points has been calculated, judge that according to its error some match points are Mismatching points.Because also participating in error, calculates Mismatching point, even a pair of correct match point, because the match point that participates in error calculating may be Mismatching point, also can cause error and the actual error of calculating inconsistent, cannot directly detect all Mismatching points according to error.Therefore, one-time detection only judges that S is Mismatching point to the match point of error maximum, and deletes this S to match point.The value minimum of S is 1, and maximal value need to be determined according to total number of match point.

(5) iterative detection stops judgement.Because one-time detection cannot detect all Mismatching points, therefore, need to complete by continuous iteration the detection of all Mismatching points.For remaining all match points, whether error in judgement is less than the threshold value of setting, if met, stops iteration, and algorithm exits.Otherwise N=N-S, returns to step (2) iterative computation again.

By above five steps, can complete the automatic detection of Mismatching point and delete Mismatching point, remaining match point is exactly the correct match point that meets error requirements.

Embodiments of the invention are realized on PC platform, and through experimental verification, method of the present invention can effectively detect and delete Mismatching point, improve the precision of images match.

It should be pointed out that the above embodiment can make the present invention of those skilled in the art's comprehend, but do not limit the present invention in any way.Therefore, it will be appreciated by those skilled in the art that still and can modify or be equal to replacement the present invention; And all do not depart from technical scheme and the improvement thereof of spirit of the present invention and technical spirit, it all should be encompassed in the middle of the protection domain of patent of the present invention.

Claims (6)

1. the Mismatching point iteration detection method based on k nearest neighbor figure, for detection of and delete the Mismatching point producing in remote sensing images couplings, improve remote sensing images matching precision, it is characterized in that comprising the steps:

(1) match point grouping, for N to be detected to match point, according to the coordinate of match point on original image, match point is divided into some groups, then each group match point is carried out to Mismatching point detection according to step (2)-(5) respectively;

(2) build k nearest neighbor figure, use the match point in original graph to build k nearest neighbor figure;

(3) calculate the error of every a pair of match point, for every a pair of match point, obtain nearest K to match point by k nearest neighbor figure, then use K to set up transformation model to match point, calculate the error of every pair of match point and transformation model as this error to match point;

(4) match point of deletion error maximum, after the error of all match points has been calculated, deletes the S of error maximum to match point;

(5) iterative detection stops judgement, if the error of all residue match points is all less than given threshold value, stops iteration; Otherwise N=N-S, returns to step (2).

2. according to the method described in claim 1, it is characterized in that: the match point grouping in step (1), the method that adopts grid to divide, first original image is carried out to grid division, sizing grid is M*N, then according to the coordinate of match point on original image, all match points are assigned to different grids; The value of M and N is determined according to original image size, ensures that each grid has up to a hundred match points.

3. according to the method described in claim 1, it is characterized in that: in step (2), the value of K is determined according to the transformation model of step (3) selection, and the value of K is greater than the needed minimum match point number of computational transformation model parameter.

4. according to the method described in claim 1, it is characterized in that: the k nearest neighbor figure in step (2) is non-directed graph, use adjacency list storage, be designated as G=(V, E), the set that wherein V is summit, the set that E is limit; Structure criterion is as follows: a summit using each match point in original image as k nearest neighbor figure, forms vertex set V=v ₁... v _n, N is total number of match point in original image; For vertex v _iif, vertex v _jv _ik nearest-neighbor in point, v _iand v _ibetween there is a nonoriented edge; Vertex v _ik nearest-neighbor in point be apart from vertex v _ia nearest K summit; Distance between two summits is the Euclidean distance of match point on original image corresponding to summit.

5. according to the method described in claim 1, it is characterized in that: the transformation model in step (3) is n rank polynomial expression, fundamental matrix, homography matrix.

6. according to the method described in claim 1, it is characterized in that: the match point error calculation formula in step (3) is:

$\mathrm{error}=\sqrt{{(x_i-x_i^{\text{'}})}^2+{(y_i-y_i^{\text{'}})}^2}$

In above formula, (x _i, y _i) be matching double points M _icoordinate on original image, (x ' _i, y ' _i) be matching double points M _icoordinate on benchmark image, is converted to the coordinate on original image according to the transformation model building.